Hi, Steve –
I’m cc-ing this to CSGnet and some other interested people, because you
have asked an important question that hasn’t really been explicitly
answered before.
SD:
At the moment, I am puzzled by your
comment about memory. You seem to be positioning reorganization as an
alternative to reinforcement-based learning theory, but learning, by most
conventional definitions, represents the creation of memories. Some set
of experiences occurs, certain elements in the brain change, and the
traces are stored in some retrievable form that makes adaptive behavior
more probable in future situations. If reorganization has no memory
element, then how can it be a learning theory? If it is not a learning
theory, why present it as an alternative to other learning theories?
Unless you are saying that learning is separate from behavior in the same
way that reorganization is separate from perceptual
control?
BP: The key here is to specify what is being learned. If it’s a route
from home to office or the spelling of a word or evaluation of a
mathematical equation, then yes, memory is involved in learning. But
“learning” is also used to mean completely different processes
that have nothing to do with the recording and playback of perceptual
information. That is why I have split reorganization off as a class of
learning that is fundamentally different from either memorizing or
applying some systematic method like long division to “learn”
what the ratio of two numbers is. I said all this, probably too tersely,
in B:CP.
Memory is involved when all that is required is to perform a specific
action or sequence of actions already in the repertoir of behaviors, like
opening a combination lock: right to 33, left two turns to 17, right to
85. Memorizing these instructions is all that is needed if you already
know how to spin the dial of of a lock, read numbers, and tell left from
right. If you don’t know how to do such things, chances are that you need
to reorganize before you can operate any combination lock no matter what
its combination is.
Reorganization alters at least connections in the nervous system and the
strengths of synapses. Hebbian learning is a simple version of this that
depends strictly on simultaneity of signals, with no other criteria for
learning. Neuroscientists have assumed some vague relationship between
Hebbian learning and Long Term Potentiation (LTP), but if you read the
textbooks carefully you find that the window of simultaneity is only
about 50 milliseconds wide for LTP to happen, and that rules out most
cases of so-called classical conditioning and Hebbian learning. I have
offered a different possible interpretation of LTP, which is that it
serves to add new copies of incoming signals that are created through
arborization of incoming axons, the signal copies thus being almost
certain to occur within 50 milliseconds of each other. This is proposed
to be a way of adjusting integer weighting of incoming signals – see
B:CP, Premises, where this method of weighting was first
described.
I usually sum this up by saying that reorganization acts on the
parameters of control: gain, delay, integration rate, and whatever else
can be adjusted by altering the way neural signals are handled. It
eventually determines HOW any behavior is carried out, and that can in
many cases determine WHAT behavior it is. In control systems, one of the
effects is to adjust the system until it controls as fast and accurately
as possible without overshoots or instability. But it can can also (in
principle, I haven’t modeled this) change what lower-order behaviors are
used to achieve higher-order ends, because in the limit, adjusting signal
weightings can amount to removing or adding connections – a weight of
zero means the signal is ignored even if it’s present. I have assumed
that reorganization can alter any physical aspect of the nervous system,
but of course have explored only a few specific instances of that idea.
None of this involves the recording and playback of perceptual
information.
Reorganization can be shown to optimize control systems starting with
either random output connections or output connections that start with
weights of zero. Once the organization of the control system has been
optimized, errors are kept small enough to make reorganization very slow
with only very small changes in parameters. There might even be a lower
threshold of error below which reorganization completely stops. The
criterion that determines when reorganization will start and how fast it
will work can be based on any variable that matters to the
organism.
The other functional architectures you sent me are a bit
hard to interpret – I can’t tell from your description whether they
would actually produce the results you say they will produce. If you have
a working model I can run on a PC I would be happy to check it out. I can
read Pascal or C source code, so if you have that available it would be
even more useful than an executable program.
You’re right, by the way, in saying I’m offering the reorganization
principle as an alternative to reinforcement theory. I don’t think
reinforcement exists – it’s a misinterpretation. And even if it did
work, it would only increase the probability of the behavioral action
that achieved a desired result, and we both know that repeating the same
action will not in general create the same result as before. Except, of
course, in a disturbance-free world.
Best regards,
Bill